[Bug] Attribution scores don't match original.

[dtch1997]

In 'notebooks/test_our_attrib_matches_original.ipynb, we check our attribution scores against those computed by the original implementation.

Annoyingly, the scores don't match. Have to figure out why this is the case...

[dtch1997]

Some quick sanity checks.

• Rerun with a single batch
• Scatterplot orig vs our edge scores
  • Make sure edge order is correct
• Check that the circuit we get is reasonable. Depends on new features

[dtch1997]

The gold standard would be to compare to Michael Hanna's implementation and ensure that we get the same scores per edge.

• This requires us to refactor his code to compute an AttributionScores object
• We also need to standardize the order of computing edges.

[dtch1997]

It's probably faster to first implement the pruning and evaluation code (which we'll need eventually anyway) and check if our circuit metrics match his.